Research On PCB Bare Board Surface Defect Detection Based On Deep Learnin

Printed circuit board(PCB)is a key component of most electronic equipment and has a very important position in the electronic information industry.As the PCB bare board continues to develop in the direction of high density,PCB manufacturers have put forward more precise requirements for the quality of PCB products.The traditional inspection methods have failed to meet the PCB inspection needs,and the detection of defects on the PCB surface in industrial production has gradually become a hot spot and a difficult research area.This paper addresses the problem of low accuracy of conventional inspection algorithms for detecting defects on the surface of bare PCBs,and studies the method of detecting defects on the surface of bare PCBs based on deep learning.The main work is as follows.(1)To address the problems of small defects and low recognition rate due to diverse defect shapes in current PCB bare board surface defect detection methods,a PCB defect detection network(DCR-FRNet)based on multi-scale feature fusion and deformable convolution is proposed.Faster RCNN is used as the benchmark network and improved and optimized in DCR-FRNet,and Res Net-101 as the base feature extraction network,constructing a multi-scale fused feature pyramid,followed by introducing deformable convolution instead of conventional convolution,and applying residual pooling to solve the model complexity problem.The experimental data show that the proposed DCRFRNet has greatly improved the detection accuracy relative to the benchmark network.(2)The noise suppression network(DF-DNet)based on multi-scale dilation convolution is proposed to address the problems that the PCB images collected in industrial production contain noise,and the existing denoising algorithms are ineffective,computationally intensive and unable to distinguish the noise features from the target features.The multi-scale module is applied in DF-DNet to extract feature maps at different levels,dilated convolution is used to capture more features,BRN is used to solve the small batch problem,and residual learning with jump connections is applied to obtain clean images.The proposed DF-DNet is compared with traditional denoising algorithms and several denoising methods for PCB noisy images in recent years for denoising experiments on PCB images to illustrate the effectiveness of the denoising effect of DF-DNet.(3)A new PCB surface defect detection method(DD-DNet)is proposed by adding a noise suppression network to DCR-FRNet to address the problem that DCR-FRNet is not effective in detecting noisy PCB bare boards.The size of PCB defects is clustered using K-means in DD-DNet,the anchor size is fine-tuned based on the clustering results and a modified loss function(DIo U)is used as the regression loss function,and the noise suppression network parameters are frozen during the training process so that the training can be completed better.The experimental results show that the proposed DD-DNet network can identify defects more effectively than the benchmark network,with a detection accuracy of 96.28% m AP,while DD-FNet can also effectively detect defects on noisy PCB bare boards,indicating that the PCB defect detection algorithm proposed in this paper is effective.